Handle missing data and outliers robustly
Overview
This question evaluates competency in machine learning preprocessing and robustness, specifically handling missingness mechanisms (MAR vs MNAR), outlier treatment, model-specific feature handling for linear and tree-based algorithms, and empirical assessment of probability calibration and interpretability.
OneMain Financial
Oct 13, 2025, 9:49 PM
Data Scientist
Technical Screen
Machine Learning
0
0
Customer Churn Modeling: Preprocessing, Missingness, Outliers, and Evaluation
Context
You are building a binary churn model for a consumer subscription/financial product. Features include:
- Numeric spend: heavy right tail with ~2% extreme outliers.
- Count variables: many zeros.
- Categorical plan types (low to moderate cardinality).
- Missing data: a mix of MAR and MNAR (e.g., some high-spend users omit income).
Answer the following:
Tasks
-
Propose end-to-end preprocessing pipelines for both:
- (A) Linear/logistic models, and
- (B) Tree ensembles (e.g., XGBoost/LightGBM/Random Forest), covering imputation (median, KNN, MICE, model-based), missingness indicators, robust scaling, and outlier treatment (winsorization vs robust estimators vs isolation-based filters).
- Explain when each choice helps or hurts and why (e.g., winsorization in logistic vs tree splits; leakage risks in MICE/KNN; effects of scaling on KNN; when to avoid isolation forest).
- Describe how you would empirically test the pipeline’s impact on probability calibration and SHAP explanations without optimistic bias.
- If ~10% of records are MNAR on a key feature, what modeling and data-collection strategies would you use to mitigate bias?
Solution
Login required
Show